23
Chenghai Yang 1 John Goolsby 1 James Everitt 1 Qian Du 2 1 USDA-ARS, Weslaco, Texas 2 Mississippi State University Applying Spectral Unmixing and Support Vector Machine to Airborne Hyperspectral Imagery for Detecting Giant Reed 1

Yang-IGARSS2011-1082.pptx

Embed Size (px)

Citation preview

Page 1: Yang-IGARSS2011-1082.pptx

Chenghai Yang1

John Goolsby1

James Everitt1

Qian Du2

1 USDA-ARS, Weslaco, Texas2 Mississippi State University

Applying Spectral Unmixing and Support Vector Machine to

Airborne Hyperspectral Imagery for Detecting Giant Reed

1

Page 2: Yang-IGARSS2011-1082.pptx

Giant Reed (Arundo donax)

Invasive weed in southern U.S. with densest stands in southern California and along the Rio Grande in Texas

Bamboo-like plant up to 10 m tall Consumes more water than native

vegetation Threat to riparian areas and watersheds Displace native vegetation, leading to the

destruction of wildlife habitats2

Page 3: Yang-IGARSS2011-1082.pptx

Mapping Invasive Weeds along the Rio Grande

Arundo

Page 4: Yang-IGARSS2011-1082.pptx

Biological Control

Difficult to control by mechanical or chemical methods

Biological control with Arundo wasps and scales Arundo wasp from Spain has been released along the

Rio Grande in Texas since July 2009 Arundo scale was released February 2011

4

Arundo scaleArundo wasp Arundo fly Arundo leafminer

Page 5: Yang-IGARSS2011-1082.pptx

5

Remote Sensing of Giant Reed

Map distribution and quantify infested areas Assess biological control efficacy Estimate water use/economic loss Necessary for its control and management Types of remote sensing imagery

– Aerial photography– Airborne multispectral imagery– Airborne hyperspectral imagery– Satellite imagery

Page 6: Yang-IGARSS2011-1082.pptx

Objectives

Evaluate linear spectral unmixing (LSU) and mixture tuned matched filtering (MTMF) for distinguishing giant reed along the Rio Grande and compare the results with those from support vector machine (SVM)

6

Page 7: Yang-IGARSS2011-1082.pptx

Study Area

7

Quemado

Page 8: Yang-IGARSS2011-1082.pptx

8

Airborne Hyperspectral Image Acquisition

Hyperspectral system Spectral range: 467–932 nm Swath width: 640 pixels Bands: 128 Radiometric: 12 bit (0–4095) Pixel size: 2.0 m

Platform Cessna 206 Altitude 2440 m & speed 180 km/h

Image date November 18, 2009 October 8, 2010

Page 9: Yang-IGARSS2011-1082.pptx

9

Normal color and CIR composites of hyperspectral image for 2009

ArundoMixed woodyMixed herbaceousBare soilWater

Normal color composite

CIR composite

Page 10: Yang-IGARSS2011-1082.pptx

10

Normal color and CIR composites of hyperspectral image for 2010

ArundoMixed woodyMixed herbaceousBare soilWater

Normal color composite

CIR composite

Page 11: Yang-IGARSS2011-1082.pptx

11

Image Correction and Rectification

Geometric correction Reference line

approach Rectification

Georeference images to UTM with GPS ground control points

102 bands were used for analysis

Raw image

Corrected image

Page 12: Yang-IGARSS2011-1082.pptx

12

Image Transformation

Minimum noise fraction (MNF) transformation was used to reduce spectral dimensionality and noise

First 30 MNF bands were selected for image classification based on eigenvalue plots and visual inspection of the MNF band images

Page 13: Yang-IGARSS2011-1082.pptx

13

Defined Classes

2009 (5 major classes) Healthy Arundo Moisture-stressed Arundo Mixed vegetation Soil/Sparse herbaceous Water

11 subclasses for classification

2010 (4 major classes) Healthy Arundo Mixed vegetation Soil/Sparse herbaceous Water

11 subclasses for classification

Page 14: Yang-IGARSS2011-1082.pptx

14

Supervised Classifications

Training samples and endmember spectra were extracted from the images for each subclass

Three classifiers were applied to 30-band MNF images Linear spectral unmixing (LSU) Mixture tuned matched filtering (MTMF) Support vector machine (SVM)

Abundance images were classified into subclasses based on maximum abundance values

Subclasses were merged into 5 major classes for 2009 and 4 classes for 2010

Page 15: Yang-IGARSS2011-1082.pptx

15

Accuracy Assessment

100 points in a stratified random pattern for the site

Error matrices for each classification Overall accuracy, producer’s accuracy,

user’s accuracy, and kappa coefficients Kappa analysis to test each classification

and the difference between any two classifications

Page 16: Yang-IGARSS2011-1082.pptx

16

Classification Maps for 2009

MTMF

SVM

CIR

Giant reedStressed giant reedMixed vegetationSoil/sparse vegetation Water

0 250 500125 m 5

Page 17: Yang-IGARSS2011-1082.pptx

17

Classification Maps for 2010

MTMF

SVM

CIR

Giant reedMixed dense vegetation Soil/Sparse vegetationWater

0 250 500125 m

Page 18: Yang-IGARSS2011-1082.pptx

18

Comparison between 2009 and 2010

MTMF-2009 SVM-2009CIR-2009

MTMF-2010 SVM-2010CIR-2010

Page 19: Yang-IGARSS2011-1082.pptx

Classifier Overallaccuracy

(%)

Overall kappa

Producer’s accuracy (PA,%) and user’s accuracy (UA,%)

Giant reed Stressed giant reed

Mixed dense vegetation

Soil/sparsevegetation

Water

PA UA PA UA PA UA PA UA PA UA

LSU

MTMF

SVM

81.0

91.0

95.0

0.757

0.885

0.936

89.3

96.4

96.4

89.3

100

100

88.2

100

100

68.2

85.0

89.5

76.0

88.0

92.0

67.9

78.6

88.5

60.0

75.0

90.0

100

100

100

100

100

100

100

100

100

Accuracy Assessment results for classification maps (2009)

19

LSU = linear spectral unmixing, MTMF = mixture tuned matched filtering, and SVM = support vector machine.

Page 20: Yang-IGARSS2011-1082.pptx

Classifier Overallaccuracy

(%)

Overall kappa

Producer’s accuracy (PA,%) and user’s accuracy (UA,%)

Giant reed Mixed densevegetation

Soil/sparsevegetation

Water

PA UA PA UA PA UA PA UA

LSU

MTMF

SVM

81.0

91.0

94.0

0.718

0.867

0.912

83.7

89.8

91.8

89.1

95.7

100

86.4

81.8

95.5

59.4

81.8

80.8

55.6

100

94.4

100

85.7

94.4

100

100

100

91.7

100

100

Accuracy Assessment results for classification maps (2010)

20

LSU = linear spectral unmixing, MTMF = mixture tuned matched filtering, and SVM = support vector machine.

Page 21: Yang-IGARSS2011-1082.pptx

Conclusions

Airborne hyperspectral imagery incorporated with image transformation and classification techniques can be a useful tool for mapping giant reed.

MTMF performed better than LSU for differentiating giant reed from associated cover types.

SVM produced the best classification results among the three classifiers examined.

Further research is needed to automate the identification of endmembers for speeding up the image classification process.

21

Page 22: Yang-IGARSS2011-1082.pptx

22

Current & Future Work

Assess the effectiveness of biological control agents (Arundo wasp and scale) with airborne imagery

Estimate ET rates of Arundo and associated vegetation

Page 23: Yang-IGARSS2011-1082.pptx

Estimating Water Use of Giant Reed Using Remote Sensing-Based Evapotranspiration Models

23

ThermalCamera

Normal Color Image Thermal Image

Arundo